A Paper Smoothing Pressure Adaptive Control Method and System Based on Intelligent Manufacturing

Abstract

This invention provides an adaptive control method and system for paper smoothing pressure based on intelligent manufacturing, belonging to the fields of process control and intelligent manufacturing. It includes: collecting data on the edge moisture content, center moisture content, ambient humidity, printing speed, drying temperature, and edge warping height of paper on a printing label production line, and obtaining the paper weight; calculating the edge-to-center moisture content gradient, ambient humidity change rate, thermal shock index, and humidity impact index; outputting the probability value of edge warping occurring within a preset time period using a pre-trained edge warping probability prediction model; and calculating a differentiated target pressure value and controlling the execution pressure when the probability value is greater than or equal to a preset probability threshold. This invention solves the problem that traditional paper smoothing control cannot predict edge warping in advance based on changes in working conditions and apply differentiated pressure to both ends of the pressure roller, leading to delayed, uneven, or damaged smoothing.

Description

Technical Field

[0001] This invention relates to the field of process control and intelligent manufacturing, specifically to an adaptive control method and system for paper smoothing pressure based on intelligent manufacturing. Background Technology

[0002] In the printing label paper production line, after the paper undergoes processes such as coating and drying, the fibers absorb moisture and expand or lose water and shrink, which easily causes regular undulating fringe curling at the edges. This can directly lead to a series of quality problems such as misregistration, die-cutting misalignment, and labels warping on finished products.

[0003] In existing technologies, the correction of paper warpage typically employs a pressure roller mechanism with fixed pressure or manual adjustment. Some improved solutions introduce single-sensor feedback, such as PID control based solely on warpage height. However, these methods cannot predict warpage in advance and can only provide passive compensation after warpage has occurred. The lagging adjustment often cannot keep up with the pace of high-speed production lines. Furthermore, they do not comprehensively consider the driving mechanism of paper deformation caused by moisture content gradients, environmental humidity impacts, and the coupling effect of thermal stress, resulting in pressure output mismatch during sudden changes in operating conditions. This can lead to either insufficient smoothing or damage to the paper surface coating due to excessive pressure.

[0004] Therefore, how to achieve forward-looking and differentiated adaptive control of smoothing pressure on the printing label paper production line is an urgent problem to be solved. Summary of the Invention

[0005] This invention provides a paper smoothing pressure adaptive control method and system based on intelligent manufacturing, which solves the problem that traditional paper smoothing control cannot predict paper edge curling in advance according to changes in working conditions and apply differentiated pressure at both ends of the pressure roller, resulting in smoothing lag, unevenness or damage to the paper.

[0006] In a first aspect, the present invention provides a paper smoothing pressure adaptive control method based on intelligent manufacturing, comprising:

[0007] Collect data on the moisture content at the edge of the paper, the moisture content in the middle of the paper, the humidity of the workshop environment, the printing speed, the drying temperature, and the warping height at the edge of the paper on the label printing production line, and obtain the paper weight of the current batch of printed label paper.

[0008] The edge-to-middle moisture content gradient is calculated based on the moisture content at the edge of the paper and the moisture content in the middle of the paper. The rate of change of ambient humidity is calculated based on the ambient humidity at multiple consecutive sampling times. The thermal shock index is calculated based on the drying temperature and the printing speed. The humidity shock index is calculated based on the rate of change of ambient humidity and the paper weight of the current batch of printed label paper.

[0009] The pre-trained frilly edge curling probability prediction model outputs the probability value of frilly edge curling within a preset time period based on the edge-middle moisture content gradient, the humidity shock index, and the thermal shock index.

[0010] When the probability value is greater than or equal to the preset probability threshold, the differential target pressure value at both ends of the smoothing roller is calculated based on the edge-middle moisture content gradient, the paper edge warping height and the paper weight, and the execution pressure of the smoothing roller is controlled according to the differential target pressure value.

[0011] Secondly, the present invention provides a paper smoothing pressure adaptive control system based on intelligent manufacturing, comprising:

[0012] The multi-source data acquisition and preprocessing module is used to collect data on the moisture content of paper edges, the moisture content of paper center, the humidity of the workshop environment, the printing speed, the drying temperature, and the warping height of paper edges on the printing label paper production line, and to obtain the paper weight of the current batch of printed label paper.

[0013] The feature index generation module is used to calculate the edge-to-middle moisture content gradient based on the moisture content of the paper edge and the moisture content of the paper middle, calculate the environmental humidity change rate based on the workshop environmental humidity at multiple consecutive sampling times, calculate the thermal shock index based on the drying temperature and the printing speed, and calculate the humidity shock index based on the environmental humidity change rate and the paper weight of the current batch of printed label paper.

[0014] The lotus leaf edge curling probability prediction module is used to output the probability value of lotus leaf edge curling within a preset time period in the future, based on the edge-middle moisture content gradient, the humidity impact index and the thermal shock index, using a pre-trained lotus leaf edge curling probability prediction model.

[0015] The differential pressure calculation and execution module is used to calculate the differential target pressure value at both ends of the smoothing pressure roller based on the edge-middle moisture content gradient, the paper edge warping height and the paper weight when the probability value is greater than or equal to a preset probability threshold, and control the execution pressure of the smoothing pressure roller according to the differential target pressure value.

[0016] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0017] The technical solution of this invention first deploys an infrared moisture meter and a laser displacement sensor between the unwinding and smoothing rollers. Combined with workshop temperature and humidity sensors, drying temperature sensors, and production system data, it simultaneously collects data on the moisture content of both sides and the middle of the paper, edge warping height, workshop ambient humidity, printing speed, drying temperature, and batch weight. This comprehensively covers the causes of edge curling from four dimensions: moisture distribution, ambient humidity, thermal stress, and physical deformation, providing a complete data foundation for subsequent analysis.

[0018] Furthermore, the raw data was processed into four physically meaningful features: edge-to-intermediate moisture content gradient, environmental humidity change rate, thermal shock index, and humidity shock index, and a time series was constructed. This transformed the original measurements, which had different dimensions and low information density, into quantitative indicators that directly characterize the curling driving force, significantly improving the input quality of subsequent prediction models.

[0019] Furthermore, a pre-trained model constructed using a three-channel LSTM parallel network is used to extract the physical driving features of the water content gradient, the hysteretic correlation features of humidity shock, and the moderating effect features of thermal shock. These features are then fused together to output the probability of leaf edge curling occurring in the future. This achieves proactive early warning, transforming traditional passive compensation into active intervention, and preventing smoothing failure caused by regulatory lag from the source.

[0020] Finally, once the predicted probability triggers the threshold, the pre-trained smoothing pressure calculation model, combined with the bias correction, outputs differentiated target pressure values ​​at both ends of the pressure roller, thereby controlling the independent output of the electric proportional valves at both ends to execute the pressure. This achieves precise pressure application to both sides of the paper, with high pressure on the severely warped side to ensure a smoothing effect, and low pressure on the slightly warped side to avoid damaging the coating.

[0021] In summary, the technical solution of this invention upgrades the management of edge curling on the printing label paper production line from passive correction relying on manual experience to proactive prevention driven by data. This allows the smoothing pressure to be applied in advance of warping formation and spatially matched to the actual state of both sides, thereby simultaneously achieving the dual goals of high smoothing rate and low damage rate in high-speed continuous production. This solves the problem that traditional paper smoothing control cannot predict edge curling in advance according to changes in working conditions and applies differentiated pressure to both ends of the pressure roller, resulting in delayed, uneven, or damaged smoothing. Attached Figure Description

[0022] Figure 1 This is a schematic flowchart of the paper smoothing pressure adaptive control method based on intelligent manufacturing provided by the present invention.

[0023] Figure 2 This is a schematic diagram of the structure of the probability prediction model for edge curling in the paper smoothing pressure adaptive control method based on intelligent manufacturing provided by the present invention.

[0024] Figure 3 This is a schematic diagram of the paper smoothing pressure adaptive control system based on intelligent manufacturing provided by the present invention.

[0025] In the attached diagram, the labels representing each component are as follows:

[0026] Multi-source data acquisition and preprocessing module 11, feature index generation module 12, leaf edge curling probability prediction module 13, differentiated pressure calculation and execution module 14. Detailed Implementation

[0027] This invention provides an adaptive control method and system for paper smoothing pressure based on intelligent manufacturing. It solves the problems of traditional paper smoothing control, which cannot predict paper edge curling in advance according to changes in working conditions and apply differentiated pressure at both ends of the pressure roller, resulting in delayed smoothing, unevenness, or damage to the paper.

[0028] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0029] Example 1, as Figure 1 As shown, this invention provides an adaptive control method for paper smoothing pressure based on intelligent manufacturing, the method comprising:

[0030] S100: Collects the moisture content at the edge of the paper, the moisture content in the middle of the paper, the humidity of the workshop environment, the printing speed, the drying temperature, and the warping height at the edge of the paper on the printing label paper production line, and obtains the paper weight of the current batch of printed label paper.

[0031] This step is used to comprehensively obtain key variables that affect paper edge curling, including uneven moisture content distribution of the paper itself, environmental humidity, thermal stress, and physical deformation that has occurred, to ensure that the complete information chain of the causes of curling is not missing.

[0032] Step S100 in the method provided by the present invention includes:

[0033] Between the unwinding station and the smoothing roller in the printing label paper production line, the moisture content of the paper edges is collected by infrared moisture meters deployed at both edges in the paper width direction, and the moisture content of the paper middle is collected by infrared moisture meters deployed in the middle of the paper width direction.

[0034] Temperature and humidity sensors are deployed in the workshop where the printing label paper production line is located to collect the ambient temperature and humidity of the workshop.

[0035] The drying temperature is collected by temperature sensors deployed at the drying station.

[0036] The paper edge warping height is collected by laser displacement sensors deployed at both edges along the width of the paper.

[0037] Get the current printing speed and the paper weight of the current batch of printed labels.

[0038] In this step, three infrared moisture meters are first arranged along the paper width direction on the conveying path between the unwinding station and the smoothing roller, respectively aimed at the left, middle, and right edge areas of the paper. The infrared moisture meters utilize the absorption characteristics of water molecules on specific wavelengths of infrared light to measure the surface moisture content of the paper in real time in a non-contact manner. The readings on the left and right sides are taken as the moisture content at the paper edge, and the reading at the middle position is taken as the moisture content at the center of the paper.

[0039] For example, on a self-adhesive label production line, the moisture content measured by the infrared moisture meter on the left is 7.8%, the moisture content measured by the infrared moisture meter in the middle is 6.2%, and the moisture content measured by the infrared moisture meter on the right is 7.5%. The moisture content of the paper edges is recorded as 7.8% on the left and 7.5% on the right, and the moisture content in the middle is 6.2%.

[0040] Secondly, temperature and humidity sensors should be installed on the walls or pillars of the workshop where the label printing production line is located, at a height of 1.5 to 2 meters above the ground, avoiding interference sources such as air conditioning vents, heating equipment, and doors and windows. The sensors should upload the workshop ambient temperature and humidity data to the control system at a fixed sampling period, such as every 10 seconds. For example, the temperature and humidity sensor might collect data showing a current workshop ambient temperature of 28.3℃ and a humidity of 52%RH.

[0041] Secondly, thermocouples or resistance temperature sensors are installed near the paper's path inside the hot air oven or infrared drying zone of the drying station to measure the actual drying temperature the paper experiences in real time, rather than the oven's set temperature, to reflect the true thermal environment. For example, using a thermocouple as the drying temperature sensor, the current drying zone temperature is measured to be 120°C.

[0042] Furthermore, a laser displacement sensor is installed above the left and right edges of the paper in the width direction. The sensor emits a laser beam to the upper surface of the paper and calculates the distance by receiving the reflected light. When the paper edge warps, the distance value deviates from the reference plane; the amount of deviation is the warping height of the paper edge. The sampling frequency is matched to the production line speed to ensure that at least 10 points are collected per meter of paper. For example, the left laser displacement sensor measures the current warping height as 1.8 mm, and the right sensor measures it as 1.2 mm.

[0043] Finally, the printing speed is directly read through communication with the printing press's main control system, taking the encoder feedback value of the main drive motor or the linear speed value converted from the inverter output frequency, reflecting the actual running speed of the paper on the production line. The basis weight of the current batch of paper is retrieved from the production management system or material information barcode, which is a batch-level static parameter, and is only updated once when changing materials for the same roll or batch of paper.

[0044] For example, the printing press main control system reports that the current printing speed is 80m / min; the production management system displays that the current batch of label paper has a basis weight of 95g / m².

[0045] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0046] In summary, this step comprehensively covers multiple factors affecting edge curling, including uneven moisture distribution, environmental humidity, printing speed, drying temperature, and edge warping height, by collecting data on both sides and the middle of the paper during a single step on the label printing production line, and simultaneously obtaining the batch paper weight. This allows subsequent analysis to address the root causes rather than just the surface appearance of warping, providing a complete and accurate data foundation for achieving forward-looking differentiated pressure control.

[0047] S200: Calculate the edge-to-middle moisture content gradient based on the moisture content at the edge of the paper and the moisture content in the middle of the paper; calculate the ambient humidity change rate based on the ambient humidity at multiple consecutive sampling times; calculate the thermal shock index based on the drying temperature and the printing speed; and calculate the humidity shock index based on the ambient humidity change rate and the paper weight of the current batch of printed label paper.

[0048] This step is used to transform raw data that is directly measured in the physical world but has low information density into quantitative indicators that can directly characterize the intensity of uneven water distribution, environmental abrupt changes, and thermal stress coupling, providing input features with clear physical meaning for subsequent prediction models.

[0049] Step S200 in the method provided by the present invention includes:

[0050] Subtracting the paper's middle moisture content from the paper's edge moisture content at the same sampling time yields the edge-middle moisture content gradient at the same sampling time.

[0051] Subtract the humidity of the workshop environment at the previous sampling time from the humidity at the current sampling time to obtain the humidity difference. Divide the humidity difference by the time interval between the current sampling time and the previous sampling time to obtain the rate of change of the ambient humidity at the current sampling time.

[0052] The thermal shock index is obtained by dividing the difference between the collected drying temperature and the workshop ambient temperature at the same sampling time by the collected printing speed.

[0053] The humidity shock index is calculated by multiplying the rate of change of ambient humidity at the current sampling time by the paper weight of the current batch of printed label paper.

[0054] In this step, at the same sampling time, the moisture content at the left edge, right edge, and center of the paper are compared to obtain the moisture content gradients at the left and right edges. The edge-center moisture content gradients directly reflect the unevenness of moisture distribution along the width of the paper. A positive value indicates that the edge is wetter than the center, and a negative value indicates that the edge is drier than the center. Both situations can cause stress differences that lead to curling.

[0055] For example, at a certain sampling time, the moisture content of the left edge is 7.8% and the moisture content of the middle is 6.2%, then the moisture content gradient of the left edge to the middle is 7.8% - 6.2% = 1.6%; the moisture content of the right edge is 7.5%, and the gradient of the right edge is 7.5% - 6.2% = 1.3%.

[0056] Next, take the ambient humidity value of the workshop at the current sampling time and subtract the ambient humidity value of the workshop at the previous sampling time to obtain the humidity difference; then divide the humidity difference by the time interval between the two sampling times to obtain the rate of change of ambient humidity at the current sampling time. A positive rate of change of ambient humidity indicates that the ambient humidity is increasing, and a negative value indicates that it is decreasing.

[0057] For example, if the current sampling time is 52%RH in the workshop, and the previous sampling time was 0.5 minutes ago, the workshop humidity was 49%RH, and the time interval is 0.5 minutes, then the rate of change of humidity is (52%-49%) / 0.5min = 6%RH / min.

[0058] Next, the difference between the drying temperature collected at the same sampling time and the ambient temperature in the workshop is calculated to obtain the temperature jump amplitude experienced by the paper as it enters the drying zone from room temperature. The difference is then divided by the current printing speed to obtain the thermal shock index. The thermal shock index characterizes the intensity of thermal shock that paper experiences per unit speed. The greater the temperature jump and the lower the printing speed, the higher the thermal shock index, indicating a more dramatic difference in thermal expansion between the paper surface and internal fibers, making it more prone to curling.

[0059] For example, if the drying temperature is measured to be 120℃, the ambient temperature in the workshop is 28℃, and the temperature difference is 92℃; and the printing speed is 80m / min, then the thermal shock index is 92℃ / 80m / min = 1.15℃·min / m.

[0060] Finally, the humidity shock index is obtained by multiplying the rate of change of ambient humidity at the current sampling time by the paper weight of the current batch of printed label paper. The humidity shock index reflects the impact intensity of changes in ambient humidity on paper of a specific weight. Under the same rate of change of humidity, the larger the paper weight, the more fibers there are, and the greater the cumulative stress generated by moisture absorption or desiccation, resulting in a higher humidity shock index.

[0061] For example, using the data from the previous example, if the rate of change in ambient humidity is 6%RH / min and the current batch of paper has a basis weight of 95g / m², then the humidity shock index is 6×95=570%RH·g / (m²·min).

[0062] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0063] In summary, this step maps the originally isolated and multidimensional raw data to the dimension of the cause of leaf edge curling, providing input with high information density and strong causal relationship for subsequent prediction models, and significantly improving the accuracy and lead time of curling prediction.

[0064] S300: Using a pre-trained frilly edge curling probability prediction model, based on the edge-middle moisture content gradient, the humidity impact index, and the thermal shock index, outputs the probability value of frilly edge curling occurring within a preset time period in the future.

[0065] like Figure 2 As shown, this step utilizes the multi-factor coupling patterns learned from historical data and uses a pre-trained fringe curling probability prediction model to make a forward-looking probability judgment before the fringe actually curls, providing a decision basis for early intervention in differentiated pressure control and avoiding passive compensation after warping has occurred.

[0066] Step S300 in the method provided by the present invention includes:

[0067] The edge-intermediate moisture content gradients of multiple consecutive sampling times before the current sampling time are used to form an edge-intermediate moisture content gradient sequence, and the humidity shock indexes of multiple consecutive sampling times before the current sampling time are used to form a humidity shock index sequence.

[0068] The edge-to-middle moisture content gradient sequence, the humidity shock index sequence, and the thermal shock index sequence are input into a pre-trained frilly edge curling probability prediction model. The physical driving channel of the frilly edge curling probability prediction model outputs a physical driving feature vector based on the edge-to-middle moisture content gradient sequence. The hysteresis correlation channel of the frilly edge curling probability prediction model outputs a hysteresis correlation feature vector based on the humidity shock index sequence. The moderating effect channel of the frilly edge curling probability prediction model outputs a moderating effect feature vector based on the thermal shock index sequence. Then, the probability fusion output layer of the frilly edge curling probability prediction model fuses the physical driving feature vector, the hysteresis correlation feature vector, and the moderating effect feature vector to output the probability value of frilly edge curling occurring within a preset time period in the future.

[0069] In this step, we first trace back from the current sampling time and take N consecutive sampling times, such as 10 consecutive sampling points, with a sampling interval of 3 seconds. This covers the left edge-middle water content gradient and the right edge-middle water content gradient recorded in a 30-second time window. These are then arranged in chronological order to form the left gradient sequence and the right gradient sequence.

[0070] The purpose of the above process is to preserve the temporal sequence information because different trends in the moisture content gradient, such as continuous expansion, stabilization, or sudden reversal, have different indicative significance for curl prediction.

[0071] For example, at the current time of the 10th second, looking back at the 10 sampling points in the past 30 seconds, the water content gradient sequence from the left edge to the middle is [1.2%, 1.3%, 1.4%, 1.5%, 1.5%, 1.6%, 1.6%, 1.7%, 1.7%, 1.6%], showing that the gradient first gradually increases and then tends to stabilize in the past 30 seconds.

[0072] Furthermore, by tracing back from the current sampling time, the humidity shock indexes calculated from N consecutive sampling times are arranged chronologically to form a humidity shock index sequence. The fluctuation pattern of the humidity shock index can reflect whether the humidity stress exerted on the paper by the workshop environment is continuous or intermittent. Continuous high shocks are more likely to cause deep moisture absorption or water loss of paper fibers than instantaneous spikes.

[0073] For example, within the same time window, the humidity shock index sequence is [520, 535, 548, 560, 568, 572, 570, 565, 558, 570], with units of %RH·g / (m²·min). This shows that the humidity shock index is generally high and fluctuates gently within the past 30 seconds, indicating that the workshop environment is in a state of continuous humidity change.

[0074] Furthermore, by tracing back from the current sampling time, the thermal shock indices calculated from N consecutive sampling times are arranged chronologically to form a thermal shock index sequence. Fluctuations in drying temperature or printing speed will be directly reflected in the temporal changes of the thermal shock index. If a sustained high value or a sudden jump appears in the sequence, it indicates that the thermal conditions of the drying station have fluctuated.

[0075] For example, within the same time window, the thermal shock index sequence is [1.10, 1.12, 1.15, 1.15, 1.13, 1.15, 1.14, 1.16, 1.16, 1.15], with units of ℃·min / m. The value fluctuates narrowly between 1.10 and 1.16, indicating that the drying conditions are generally stable.

[0076] In this step, the pre-training process of the ruffle edge curling probability prediction model includes:

[0077] Obtain a historical production sample set. Each historical production sample in the historical production sample set contains an edge-intermediate moisture content gradient sequence, a humidity shock index sequence, and a thermal shock index sequence for multiple consecutive sampling times in a historical production batch, as well as a corresponding fringe curling occurrence label. The fringe curling occurrence label indicates whether fringe curling actually occurred in the future preset time period in the historical production batch.

[0078] An initial fringe edge curling probability prediction model is constructed, which includes a physical driving channel, a hysteresis correlation channel, a moderating effect channel, and a probability fusion output layer.

[0079] The edge-to-intermediate moisture content gradient sequence, humidity shock index sequence, and thermal shock index sequence from the historical production sample set are used as input features. The corresponding frilly edge curling occurrence labels are used as output labels. The training objective is to minimize the binary cross-entropy loss function. The initial frilly edge curling probability prediction model is trained in a supervised manner until it is verified to converge, thus obtaining the trained frilly edge curling probability prediction model. The binary cross-entropy loss function calculates the error between the probability value output by the initial frilly edge curling probability prediction model and the frilly edge curling occurrence label.

[0080] Specifically, firstly, operational data from multiple historical production batches are extracted from the production line's historical operation records, following the same data structure and sampling rules as real-time prediction. Each historical production sample contains three components: an edge-to-intermediate moisture content gradient sequence across multiple consecutive sampling times within a certain time period; a humidity shock index sequence within the same time period; and a thermal shock index sequence within the same time period. The construction method for these three sets of sequences is completely consistent with that of the real-time inference stage, ensuring alignment between the distribution of training and inference data.

[0081] For example, 5,000 valid historical samples were extracted from production records over the past three months. Sample No. 372 was taken from the production period of a batch of 80g / m² coated paper labels. Its moisture content gradient sequence from the left edge to the center is [1.1%, 1.3%, 1.4%, 1.6%, 1.8%], its humidity shock index sequence is [480, 495, 510, 520, 535], and its thermal shock index sequence is [0.95, 0.98, 1.02, 1.05, 1.08]. Only the first five data points from these sequences are used as examples.

[0082] Furthermore, for each historical production sample, through manual visual inspection or records from the end-of-line quality inspection system, it is confirmed whether the paper actually exhibits edge curling within a preset future time period following the sample's current time period. The preset future time period is consistent with the real-time predicted preset time period, such as 30 seconds later. If the paper actually exhibits edge curling, the label is marked as 1; if not, the label is marked as 0. The accuracy of the labels directly affects the model's predictive ability; therefore, it requires dual confirmation through quality inspection records and image re-inspection.

[0083] For example, according to the quality inspection records, within 30 seconds after the time period corresponding to sample number 372, a noticeable curling of the left edge of the paper was detected at the end of the production line, and the label for the curling of the sample's edge was marked as 1. However, another sample from the same batch had normal paper condition in the following 30 seconds, and its label was marked as 0.

[0084] Furthermore, an initial prediction model for the probability of edge curling of the ruffled leaf is constructed, including:

[0085] A first long short-term memory neural network is used to construct a physical driving channel. The physical driving channel takes the edge-middle water content gradient sequence as input, extracts the water content gradient temporal cumulative features, and outputs a physical driving feature vector.

[0086] A second long short-term memory neural network is used to construct a hysteresis correlation channel. The hysteresis correlation channel takes the humidity shock index sequence as input, extracts the hysteresis response features from humidity shock to moisture content gradient change, and outputs a hysteresis correlation feature vector.

[0087] A third long short-term memory neural network is used to construct a regulation effect channel. The regulation effect channel takes the thermal shock index sequence as input, extracts the inverse regulation features of thermal shock on water content gradient, and outputs a regulation effect feature vector.

[0088] A probabilistic fusion output layer is constructed. The probabilistic fusion output layer concatenates the physical driving feature vector, the hysteresis correlation feature vector, and the moderating effect feature vector in the feature dimension to obtain a concatenated fusion vector. The concatenated fusion vector is mapped to a one-dimensional feature value through a fully connected layer. Then, the one-dimensional feature value is mapped to a probability value of the ruffled edge curling occurring in a future preset time period through a Sigmoid activation function.

[0089] The physical driving channel, the hysteresis correlation channel, the modulation effect channel, and the probability fusion output layer are connected to form an initial lotus leaf edge curling probability prediction model.

[0090] Specifically, a physical driving channel is first constructed using a first long short-term memory neural network, taking the edge-to-intermediate moisture content gradient sequence as input. The gating mechanism of the long short-term memory neural network can capture long-term dependencies in the sequence, which is used here to extract the cumulative change features of the moisture content gradient over time, such as whether the gradient continues to expand, whether it tends to plateau, or whether it reverses. The output of the physical driving channel is a physical driving feature vector, which characterizes the stress accumulation process caused by uneven moisture distribution within the paper fibers.

[0091] Among them, the Long Short-Term Memory (LSTM) network is a recurrent neural network structure with memory units. Through input gates, forget gates, and output gates, it controls the flow and storage of information, effectively capturing long-term dependencies in time-series data and avoiding the gradient vanishing problem that traditional recurrent neural networks encounter when processing long sequences. In this invention, three independent LSTM networks are deployed for each of the three channels, each learning the temporal patterns of its corresponding feature sequences.

[0092] For example, if the water content gradient sequence from the left edge to the middle [1.2%, 1.3%, 1.4%, 1.5%, 1.5%, 1.6%, 1.6%, 1.7%, 1.7%, 1.6%] is input into the physical driving channel, the LSTM network extracts the trend feature of the sequence first rising and then flattening out, and outputs a physical driving feature vector with a dimension of 16.

[0093] Furthermore, a second long short-term memory neural network is used to construct a hysteresis correlation channel, with the humidity shock index sequence as input. The effect of environmental humidity changes on paper moisture content is not instantaneous, but involves a diffusion and equilibrium hysteresis process. That is, after the peak of the humidity shock, the internal moisture content gradient of the paper may take several seconds to tens of seconds to reach its maximum response. The hysteresis correlation channel specifically learns the above-mentioned hysteresis response pattern and outputs a hysteresis correlation feature vector to characterize the delayed effect of humidity shock on the time axis.

[0094] For example, if the humidity shock index sequence [520,535,548,560,568,572,570,565,558,570] is input into the hysteresis correlation channel, the LSTM network will capture the slight decline features of the sequence after a sustained high shock and output a hysteresis correlation feature vector with a dimension of 16.

[0095] Furthermore, a third long short-term memory neural network is used to construct a modulation effect channel, with the thermal shock index sequence as input. Thermal shock during the drying process has a reverse modulation effect on the paper moisture content gradient—high temperatures accelerate the evaporation of moisture at the edges, thereby altering the original gradient distribution. The modulation effect channel extracts this reverse modulation temporal pattern contained in the thermal shock index sequence and outputs a modulation effect feature vector, which is used to characterize the corrective or aggravating effect of thermal stress on curl formation.

[0096] For example, by inputting the thermal shock index sequence [1.10,1.12,1.15,1.15,1.13,1.15,1.14,1.16,1.16,1.15] into the moderating effect channel, the LSTM network identifies the pattern of stable fluctuation of the thermal shock index within a narrow range and outputs a moderating effect feature vector with a dimension of 16.

[0097] Furthermore, the physical driving feature vector, hysteresis correlation feature vector, and moderating effect feature vector output from the first three steps are concatenated along their feature dimensions to obtain a concatenated fusion vector. Each of the three channels outputs a feature vector with a dimension of 16, which is concatenated to obtain a fusion vector with a dimension of 48. Subsequently, the fusion vector is mapped to one-dimensional feature values ​​through a fully connected layer, and then the one-dimensional feature values ​​are mapped to the interval between 0 and 1 through a Sigmoid activation function, which represents the probability value of the ruffled edge curling occurring within a preset time period in the future.

[0098] Among them, Sigmoid is a non-linear function that maps any real number to the interval (0,1). It is the most commonly used activation function for output layers in binary classification probability output scenarios, and the output value can be directly interpreted as the probability of an event occurring.

[0099] For example, after concatenating the three feature vectors into a 48-dimensional fusion vector, it is mapped to a one-dimensional value of 1.27 through a fully connected layer, and then outputs a probability value of 0.78 through the Sigmoid function, which means that the probability of the ruffled edge curling within the next 30 seconds is 78%.

[0100] Finally, the physical driving channel, hysteresis correlation channel, moderating effect channel, and probabilistic fusion output layer are connected sequentially according to the data flow. Each of the three channels receives its respective feature sequence input in parallel, independently extracts temporal features, and then merges them into the probabilistic fusion output layer for fusion and probability mapping, forming the initial bobbin curl probability prediction model. The initial network weights are randomly initialized and need to be trained using historical production sample sets in subsequent steps before deployment.

[0101] Furthermore, the probability values ​​output by the initial ruffle edge curling probability prediction model and their corresponding true labels are substituted into the binary cross-entropy loss function to calculate the prediction error for the current batch. The binary cross-entropy loss function is: -[true label · log(model output prediction probability) + (1 - true label) · log(1 - model output prediction probability)]. When the true label is 1, the loss is -log(model output prediction probability), and the closer the prediction probability is to 1, the smaller the loss; when the true label is 0, the loss is -log(1 - model output prediction probability), and the closer the prediction probability is to 0, the smaller the loss. All the above logarithmic calculation results are base e.

[0102] Among them, the binary cross-entropy loss function is the most commonly used loss function in binary classification tasks, measuring the difference between the probability distribution predicted by the model and the true label distribution. Its gradient tends to be flat when the predicted value is close to the true value, and large when the predicted value deviates significantly from the true value, exhibiting good convergence characteristics.

[0103] For example, if sample number 372 has a true label of 1 and a predicted probability of 0.62, then the loss value is -log(0.62) ≈ 0.478. Another sample in the same batch with a label of 0 has a predicted probability of 0.35, and its loss value is -log(1-0.35) = -log(0.65) ≈ 0.430. The total loss for this batch is the mean of the losses for all samples.

[0104] Furthermore, based on the binary cross-entropy loss value, the gradient of the loss with respect to the weights and biases of each layer in the model is calculated using the backpropagation algorithm, including the gating parameters within the three LSTM channels and the weight matrix of the fully connected layers. Subsequently, a gradient descent optimizer is used to update the model parameters along the negative gradient direction, making the model's predictions in the next iteration closer to the true labels. The learning rate controls the step size of each parameter update and must be set to an appropriate value to avoid slow convergence or oscillations.

[0105] For example, the batch loss in this training round is 0.45. After the optimizer calculates the gradient of each layer, it fine-tunes the weight matrix of the LSTM forget gate in the physical driving channel from the initial value by 0.002, and updates a certain weight of the fully connected layer from 0.15 to 0.148, thus completing a parameter update.

[0106] Finally, repeat the aforementioned steps, iterating through all samples in the historical production sample set to complete one training epoch. After several epochs, evaluate the model performance on an independent validation set, recording the validation loss and prediction accuracy. When the validation loss no longer decreases over several consecutive epochs, or begins to show an upward trend, the model is considered to have converged, training is stopped, and the current model parameters are saved as the trained ruffle probability prediction model.

[0107] For example, when training reached the 48th round, the validation loss decreased from the initial 0.68 to 0.22 and did not decrease for 5 consecutive rounds. The validation set accuracy reached 91.5% and the recall rate was 87.2%, which met the preset convergence condition, and the training terminated.

[0108] Among them, the pre-trained fringe curling probability prediction model refers to a model that has completed all training, validation, and parameter solidification using historical production data or historical control records in an offline environment before being put into real-time operation on the production line, and has the ability to directly infer. After training convergence, it is deployed to the edge computing device of the production line, and only performs forward inference calculations during real-time operation, without updating parameters.

[0109] In the above-mentioned probability prediction model for lotus leaf edge curling, the configuration of the three LSTM channels is as follows: The physical driving channel uses a two-layer LSTM with a hidden layer dimension of 32, a time step of 10 steps, covering 30 seconds, and an inter-layer dropout of 0.2. Finally, the hidden state at the last time step is mapped to a 16-dimensional feature vector through a linear layer. The hysteresis correlation channel also uses a two-layer LSTM with a hidden layer dimension of 32, but the time step is extended to 15 steps, covering 45 seconds, to capture the longer hysteresis response of humidity impact on moisture content changes. The moderating effect channel, due to the lower temporal complexity of the thermal shock index, uses a single-layer LSTM with a hidden layer dimension of 16, a time step of 10 steps, and the hidden state is directly output as a 16-dimensional feature vector. The 16-dimensional feature vectors output by the three channels are concatenated into a 48-dimensional fusion vector, which is then compressed to 24 dimensions through a fully connected layer, activated by ReLU, with a dropout of 0.3, and then mapped to 1 dimension. Finally, the probability value is output through the Sigmoid function. The Adam optimizer was used during the training phase, with an initial learning rate of 0.001 that decayed by half every 20 rounds and a batch size of 64.

[0110] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0111] In summary, this step enables the production line to receive an early warning signal before the warpage defect actually forms, fundamentally avoiding smoothing failure caused by adjustment lag, and providing accurate triggering timing and decision-making basis for subsequent differentiated pressure control.

[0112] S400: When the probability value is greater than or equal to the preset probability threshold, calculate the differentiated target pressure value at both ends of the smoothing roller based on the edge-middle moisture content gradient, the paper edge warping height and the paper weight, and control the execution pressure of the smoothing roller according to the differentiated target pressure value.

[0113] This step calculates target pressure values ​​for both ends of the smoothing roller based on the moisture content gradient, warpage height, and paper basis weight when the predicted probability value is greater than or equal to the preset probability threshold, and performs differentiated control accordingly. This avoids the problems of insufficient pressure on one side leading to incomplete smoothing and excessive pressure on the other side damaging the paper, which are common in traditional equal pressure control, thus achieving precise and low-loss paper smoothing.

[0114] Step S400 in the method provided by the present invention includes:

[0115] The probability value is compared with the preset probability threshold. When the probability value is greater than or equal to the preset probability threshold, the edge-to-middle moisture content gradient, the paper edge warping height and the paper weight of the current batch of printed label paper are input into the pre-trained smoothing pressure calculation model. The smoothing pressure calculation model outputs the predicted differential target pressure value.

[0116] The bias correction amount is determined based on the historical prediction error direction distribution of the smoothing pressure calculation model on the validation set;

[0117] The predicted differential target pressure value is added to the bias correction amount to obtain the differential target pressure value at both ends of the smoothing roller;

[0118] According to the differentiated target pressure value, the output pressure of the electric proportional valves of the cylinders at both ends of the smoothing roller is controlled respectively, so that different execution pressures are applied to the two ends of the smoothing roller.

[0119] In this step, the probability value output by the pre-trained ruffle edge curling probability prediction model is first compared with a preset probability threshold. The preset probability threshold is set based on the production line's tolerance for false negatives and false positives, typically ranging from 0.55 to 0.75. When the probability value is greater than or equal to the preset probability threshold, the risk of ruffle edge curling is determined to have entered the range requiring proactive intervention, triggering a differentiated pressure control process; when the probability value is less than the threshold, the current basic pressure is maintained, and monitoring continues.

[0120] Furthermore, when the probability value is greater than or equal to a preset probability threshold, the following parameters will be combined into a set of input features: the moisture content gradient between the left edge and the middle, the moisture content gradient between the right edge and the middle, the warp height of the left paper edge, the warp height of the right paper edge, and the paper weight of the current batch of printed label paper. These features will then be input into the pre-trained smoothing pressure calculation model. The paper weight is a batch-level static parameter, and its value is directly retrieved from the current batch at the triggering time.

[0121] For example, the probability prediction model for edge curling outputs a current probability value of 0.78, while the preset probability threshold is 0.70. Since 0.78 is greater than 0.70, intervention is triggered. At the trigger point, the moisture content gradient from the left edge to the middle is 1.6%, and the gradient on the right side is 1.3%; the warping height of the left paper edge is 1.8 mm, and the warping height on the right side is 1.2 mm; the current batch paper weight is 95 g / m². These five parameters are combined into an input feature vector, which is then input into a pre-trained smoothing pressure calculation model.

[0122] In this step, the pre-training process of the stress relief calculation model includes:

[0123] Obtain a historical control sample set. Each historical control sample in the historical control sample set contains an edge-to-middle moisture content gradient, paper edge warping height, and paper weight in a historical edge curling control event, as well as a corresponding optimal differentiated target pressure value label.

[0124] The historical regulation sample set is augmented to obtain an augmented historical regulation sample set.

[0125] An initial stress calculation model was constructed based on a fully connected neural network.

[0126] Using the edge-to-middle moisture content gradient, paper edge warping height, and paper basis weight in the enhanced historical control sample set as input features, and the corresponding optimal differentiated target pressure value label as output label, the initial smoothing pressure calculation model is trained in a supervised manner until the verification convergence is obtained, thus obtaining the trained smoothing pressure calculation model.

[0127] Specifically, the historical regulation sample set is augmented to obtain an augmented historical regulation sample set, which includes:

[0128] For each historical control sample in the historical control sample set, obtain the edge-to-middle moisture content gradient and the paper edge warping height of the historical control sample;

[0129] Within a preset perturbation range, a first random micro-perturbation is applied to the edge-middle moisture content gradient of the historical control sample, and a second random micro-perturbation is applied to the paper edge warping height of the historical control sample, thereby generating the perturbed edge-middle moisture content gradient and the perturbed paper edge warping height.

[0130] The perturbed edge-to-middle moisture content gradient and the perturbed paper edge warp height are combined with the paper weight of the historical control sample to form the input features of the enhanced sample. The optimal differentiated target pressure value label corresponding to the historical control sample is used as the output label of the enhanced sample.

[0131] The enhanced samples generated from all historical control samples are merged with the historical control sample set to obtain the enhanced historical control sample set.

[0132] Specifically, data for each edge curling control event is first extracted from the production line's historical control records. Each historical control sample includes the edge-to-center moisture content gradient, paper edge warping height, and paper basis weight at the time the control event occurred, as well as an output label. The output label is confirmed by experienced operators after multiple trials, representing the optimal differentiated target pressure value that precisely smooths the warping without damaging the paper. The label must include two pressure values: one for the left end and one for the right end of the smoothing roller.

[0133] For example, 300 valid historical control samples were extracted from the control records of the past six months. Among them, sample No. 85 recorded a control event: the moisture content gradient from the left edge to the middle was 1.5%, the warping height of the left paper edge was 1.6 mm, the paper weight was 80 g / m², and the operator finally determined that the optimal pressure on the left was 85 N and the optimal pressure on the right was 62 N, which is the output label No. 85.

[0134] Secondly, since the number of actual control events triggered on the production line is limited, directly using the original sample set for training may lead to insufficient model generalization ability. Therefore, a small perturbation is applied to each historical control sample to expand the sample size.

[0135] Specifically, a preset perturbation range is first set for each historical control sample. The rules for setting the preset perturbation range are as follows: the perturbation amplitude of the edge-to-center moisture content gradient does not exceed ±10% of the original value, and the absolute value does not exceed 0.3 percentage points; the perturbation amplitude of the paper edge warping height does not exceed ±0.3mm of the original value. Small perturbations simulate the tiny random fluctuations in sensor measurement noise and paper condition, while excessive perturbations may deviate from the actual physical laws.

[0136] Specifically, within the perturbation range, a first random micro-perturbation is applied to the edge-to-center moisture content gradient, and a second random micro-perturbation is applied to the paper edge warp height, generating perturbed feature values. The perturbation method involves adding a random offset that is uniformly or normally distributed within a preset range to the original value. The paper basis weight, as a batch static parameter, is not perturbed. The perturbed feature values, combined with the original paper basis weight, constitute the input features of the enhanced sample, and the optimal differentiated target pressure value label of the original sample is directly used as the output label of the enhanced sample.

[0137] For example, the perturbation range of the original gradient of sample No. 85 (1.5%) is ±0.15%, and the perturbation range of the original warp height of 1.6 mm is ±0.3 mm. When generating enhanced samples, a random perturbation of +0.08% is applied to the gradient, resulting in a perturbed gradient of 1.58%; a random perturbation of -0.12 mm is applied to the warp height, resulting in a perturbed warp height of 1.48 mm. This is combined with a paper weight of 80 g / m² as the input feature for the enhanced samples, and the output labels remain 85N on the left and 62N on the right. This operation is repeated multiple times for sample No. 85, with different random perturbations each time, generating multiple enhanced samples.

[0138] Next, the aforementioned perturbation enhancement operation is performed on each sample in the historical control sample set, and multiple enhanced samples can be generated from each original sample. All enhanced samples are merged with the original historical control sample set to obtain an enhanced historical control sample set with a significantly expanded sample size.

[0139] For example, if the original sample set has 300 samples, and 10 augmented samples are generated for each sample, a total of 3,000 augmented samples are obtained. After merging with the original 300 samples, the augmented sample set has a total of 3,300 samples.

[0140] Furthermore, an initial smoothing pressure calculation model was constructed based on a fully connected neural network. The input layer of the smoothing pressure calculation model receives five neurons, corresponding to the moisture content gradient between the left and right edges, the warp height of the left and right paper edges, and the paper weight, respectively. The middle layer contains several hidden layers, each employing a ReLU activation function to introduce nonlinear mapping capabilities. The output layer has two neurons, outputting the predicted differential target pressure values ​​for the left and right ends of the smoothing roller, respectively. Because the pressure value is a continuous regressor, the output layer either does not use an activation function or uses a linear activation function, and the value range is unrestricted.

[0141] The ReLU activation function is a function that sets input values ​​less than zero to zero and keeps input values ​​greater than zero unchanged. It can effectively alleviate the gradient vanishing problem and is a commonly used activation function for hidden layers in deep learning.

[0142] For example, a fully connected network with 3 hidden layers and 64 neurons in each layer is constructed, with 5 neurons in the input layer and 2 neurons in the output layer, and a total of about 8,000 parameters, as the initial stress relief calculation model.

[0143] Finally, the input features from the enhanced historical regulation sample set are used as the model input, and the corresponding optimal differentiated target pressure value label is used as the model output target. The mean squared error loss function is used to measure the deviation between the predicted pressure value and the labeled pressure value, and the network weights are updated through backpropagation and gradient descent optimizer. During training, the loss value is monitored on the validation set. When the validation loss no longer decreases for several consecutive rounds, the model is considered to have converged, and the model parameters are saved.

[0144] For example, using 70% of 3300 samples as the training set, 15% as the validation set, and 15% as the test set, the mean squared error loss function was trained to a stable validation loss of around 4.2 after 65 rounds and no longer decreased. The average prediction error on the test set was ±3.5N, which met the production line accuracy requirements, and the training was completed.

[0145] Among them, the pre-trained stress-relief calculation model refers to a model that has completed all training, verification, and parameter solidification using historical production data or historical control records in an offline environment before being put into real-time operation on the production line, and has the ability to perform direct inference. After training convergence, it is deployed to the edge computing device of the production line, and only performs forward inference calculations during real-time operation, without updating parameters.

[0146] In this step, the bias correction amount is determined based on the historical prediction error direction distribution of the smoothing pressure calculation model on the validation set, including:

[0147] Construct a validation set containing multiple validation samples, where each validation sample includes the edge-to-center moisture content gradient, paper edge warping height, paper basis weight, and the corresponding optimal differentiated target pressure value label;

[0148] For each validation sample in the validation set, the edge-to-center moisture content gradient of the validation sample, the paper edge warping height of the validation sample, and the paper weight are input into the trained smoothing pressure calculation model, and the predicted differential target pressure value of the validation sample is output.

[0149] For each validation sample in the validation set, the optimal differentiated target pressure value label of the validation sample is calculated, and the predicted differentiated target pressure value of the validation sample is subtracted to obtain the prediction error value of the validation sample.

[0150] The arithmetic mean of the prediction error values ​​of all validation samples in the validation set is used as the bias correction.

[0151] Specifically, during the training of the pressure calculation model, a portion of the enhanced historical control sample set is proportionally allocated as a validation set. Each sample in the validation set also contains complete input features and output labels: the input features are the moisture content gradient between the left edge and the middle, the moisture content gradient between the right edge and the middle, the warp height of the left paper edge, the warp height of the right paper edge, and the paper weight; the output labels are the corresponding optimal pressure values ​​at the left and right ends. The validation set does not participate in model training; it is only used to evaluate model performance and calculate bias corrections.

[0152] 15%, or 495 samples, were randomly selected from the enhanced 3300 samples to form the validation set. The input features of validation sample No. 126 were: left gradient 1.8%, right gradient 1.1%, left warp height 2.0 mm, right warp height 0.9 mm, and paper weight 90 g / m²; the output labels were: left optimal pressure 102 N and right optimal pressure 58 N.

[0153] Furthermore, the input features of each sample in the validation set are fed into the trained smoothing pressure calculation model. After forward propagation through a fully connected network, the output layer generates two values, which correspond to the predicted differential target pressure values ​​at the left and right ends of the smoothing roller, respectively.

[0154] For example, when the input features of validation sample number 126 are input into the trained model, the model outputs a prediction pressure of 96.5N on the left and 62.3N on the right. However, the sample's label values ​​are 102N on the left and 58N on the right, indicating that the prediction on the left is too low and the prediction on the right is too high.

[0155] Furthermore, for each validation sample, the predicted differentiated target pressure value output by the model is subtracted from the optimal differentiated target pressure value label to obtain the prediction error values ​​on the left and right sides of the sample. A positive error value indicates that the model predicts too low, and a negative error value indicates that the model predicts too high.

[0156] For example, the error value on the left side of the 126th validation sample is 102N - 96.5N = +5.5N, which means that the left side prediction is 5.5N lower; the error value on the right side is 58N - 62.3N = -4.3N, which means that the right side prediction is 4.3N higher.

[0157] Finally, the left-hand bias correction is obtained by taking the arithmetic mean of the prediction error values ​​on the left side of all validation samples in the validation set; the right-hand bias correction is obtained by taking the arithmetic mean of the prediction error values ​​on the right side of all validation samples. The bias correction reflects the direction and magnitude of the systematic bias of the model on the validation set. A positive value indicates that the overall prediction of the model is too low and needs to be corrected upward, while a negative value indicates that the overall prediction of the model is too high and needs to be corrected downward.

[0158] For example, the arithmetic mean of the left-hand error values ​​of 495 validation samples is +3.8N, and the arithmetic mean of the right-hand error values ​​is -2.4N. Therefore, the left-hand bias correction is +3.8N, and the right-hand bias correction is -2.4N. These two bias corrections will be fixed after model deployment. During each real-time inference, the original model output will be added to the corresponding bias correction to obtain the final differentiated target pressure value.

[0159] In this step, after obtaining the bias correction, the left-end predicted differential target pressure value output by the pre-trained pressure smoothing calculation model is added to the left-end bias correction to obtain the final left-end differential target pressure value; the right-end predicted differential target pressure value is added to the right-end bias correction to obtain the final right-end differential target pressure value.

[0160] For example, the pressure calculation model outputs a predicted pressure of 92.5 N on the left and 65.0 N on the right. The left-side bias correction is +3.8 N, and the right-side bias correction is -2.4 N. After correction, the final target pressure on the left is 92.5 + 3.8 = 96.3 N, and the final target pressure on the right is 65.0 + (-2.4) = 62.6 N.

[0161] Finally, each end of the smoothing roller is equipped with an independent cylinder and an electro-proportional valve. The electro-proportional valve receives the target pressure signal sent by the control system, adjusts the valve core opening through the built-in PID controller, and precisely adjusts the air source pressure to the target value before outputting it to the cylinder. The left electro-proportional valve outputs the air pressure value corresponding to 96.3N, and the right electro-proportional valve outputs the air pressure value corresponding to 62.6N, so that the left side of the smoothing roller applies a downward pressure of approximately 96.3N, and the right side applies a downward pressure of approximately 62.6N. The pressure difference between the two ends allows the more warped left side to receive a stronger smoothing force, while the less warped right side receives a moderate smoothing force, achieving differentiated control.

[0162] For example, the controller converts the target pressure of 96.3N on the left into a control signal for the electro-proportional valve, such as the corresponding value in a 0-10V voltage signal or a 4-20mA current signal, and sends it to the electro-proportional valve on the left. The valve body then precisely adjusts the 0.6MPa air source to the corresponding cylinder output pressure of 96.3N; similarly, the right side is adjusted to the corresponding pressure of 62.6N. The smoothing roller applies differentiated pressure to the paper surface with a higher pressure on the left and a lower pressure on the right. The high warp area on the left is effectively flattened, while the right side is sufficiently smoothed without excessive pressure that could damage the coating.

[0163] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0164] In summary, this step achieves precise pressure application and differentiated control based on demand, fundamentally resolving the contradiction in traditional equal pressure smoothing where insufficient pressure on one side leads to warping residue, while excessive pressure on the other side causes paper indentations or coating damage.

[0165] In summary, the management of edge curling on the label printing production line has been upgraded from passive correction relying on manual experience to proactive prevention driven by data. This allows the smoothing pressure to precede the formation of warping in time and to match the actual state of both sides in space, thereby achieving the dual goals of high smoothing rate and low damage rate simultaneously in high-speed continuous production.

[0166] Example 2, as Figure 3 As shown, this invention provides an adaptive control system for paper smoothing pressure based on intelligent manufacturing, the system comprising:

[0167] The multi-source data acquisition and preprocessing module 11 is used to collect the paper edge moisture content, paper middle moisture content, workshop ambient humidity, printing speed, drying temperature and paper edge warping height on the printing label paper production line, and to obtain the paper weight of the current batch of printed label paper.

[0168] This includes collecting data on the moisture content at the edges of the printed label paper, the moisture content in the middle of the paper, the ambient humidity in the workshop, the printing speed, the drying temperature, and the edge warping height of the paper on the printing label paper production line, and obtaining the paper weight of the current batch of printed label paper, including:

[0169] Between the unwinding station and the smoothing roller in the printing label paper production line, the moisture content of the paper edges is collected by infrared moisture meters deployed at both edges in the paper width direction, and the moisture content of the paper middle is collected by infrared moisture meters deployed in the middle of the paper width direction.

[0170] Temperature and humidity sensors are deployed in the workshop where the printing label paper production line is located to collect the ambient temperature and humidity of the workshop.

[0171] The drying temperature is collected by temperature sensors deployed at the drying station.

[0172] The paper edge warping height is collected by laser displacement sensors deployed at both edges along the width of the paper.

[0173] Get the current printing speed and the paper weight of the current batch of printed labels.

[0174] The feature index generation module 12 is used to calculate the edge-to-middle moisture content gradient based on the moisture content at the edge of the paper and the moisture content in the middle of the paper, calculate the environmental humidity change rate based on the workshop environmental humidity at multiple consecutive sampling times, calculate the thermal shock index based on the drying temperature and the printing speed, and calculate the humidity shock index based on the environmental humidity change rate and the paper weight of the current batch of printed label paper.

[0175] The process includes calculating the edge-to-center moisture content gradient based on the paper edge moisture content and the paper center moisture content; calculating the environmental humidity change rate based on the workshop environmental humidity at multiple consecutive sampling times; calculating the thermal shock index based on the drying temperature and the printing speed; and calculating the humidity shock index based on the environmental humidity change rate and the paper weight of the current batch of printed label paper.

[0176] Subtracting the paper's middle moisture content from the paper's edge moisture content at the same sampling time yields the edge-middle moisture content gradient at the same sampling time.

[0177] Subtract the humidity of the workshop environment at the previous sampling time from the humidity at the current sampling time to obtain the humidity difference. Divide the humidity difference by the time interval between the current sampling time and the previous sampling time to obtain the rate of change of the ambient humidity at the current sampling time.

[0178] The thermal shock index is obtained by dividing the difference between the collected drying temperature and the workshop ambient temperature at the same sampling time by the collected printing speed.

[0179] The humidity shock index is calculated by multiplying the rate of change of ambient humidity at the current sampling time by the paper weight of the current batch of printed label paper.

[0180] The 13th lint curling probability prediction module is used to output the probability value of lint curling occurring within a preset time period in the future, based on the edge-middle moisture content gradient, the humidity impact index and the thermal shock index, using a pre-trained lint curling probability prediction model.

[0181] Among them, the pre-trained lotus leaf edge curling probability prediction model, based on the edge-to-center moisture content gradient, the humidity shock index, and the thermal shock index, outputs the probability value of lotus leaf edge curling occurring within a preset time period in the future, including:

[0182] The edge-intermediate moisture content gradients of multiple consecutive sampling times before the current sampling time are used to form an edge-intermediate moisture content gradient sequence, and the humidity shock indexes of multiple consecutive sampling times before the current sampling time are used to form a humidity shock index sequence.

[0183] The edge-to-middle moisture content gradient sequence, the humidity shock index sequence, and the thermal shock index sequence are input into a pre-trained frilly edge curling probability prediction model. The physical driving channel of the frilly edge curling probability prediction model outputs a physical driving feature vector based on the edge-to-middle moisture content gradient sequence. The hysteresis correlation channel of the frilly edge curling probability prediction model outputs a hysteresis correlation feature vector based on the humidity shock index sequence. The moderating effect channel of the frilly edge curling probability prediction model outputs a moderating effect feature vector based on the thermal shock index sequence. Then, the probability fusion output layer of the frilly edge curling probability prediction model fuses the physical driving feature vector, the hysteresis correlation feature vector, and the moderating effect feature vector to output the probability value of frilly edge curling occurring within a preset time period in the future.

[0184] The pre-training process of the ruffle edge curling probability prediction model includes:

[0185] Obtain a historical production sample set. Each historical production sample in the historical production sample set contains an edge-intermediate moisture content gradient sequence, a humidity shock index sequence, and a thermal shock index sequence for multiple consecutive sampling times in a historical production batch, as well as a corresponding fringe curling occurrence label. The fringe curling occurrence label indicates whether fringe curling actually occurred in the future preset time period in the historical production batch.

[0186] An initial fringe edge curling probability prediction model is constructed, which includes a physical driving channel, a hysteresis correlation channel, a moderating effect channel, and a probability fusion output layer.

[0187] The edge-to-intermediate moisture content gradient sequence, humidity shock index sequence, and thermal shock index sequence from the historical production sample set are used as input features. The corresponding frilly edge curling occurrence labels are used as output labels. The training objective is to minimize the binary cross-entropy loss function. The initial frilly edge curling probability prediction model is trained in a supervised manner until it is verified to converge, thus obtaining the trained frilly edge curling probability prediction model. The binary cross-entropy loss function calculates the error between the probability value output by the initial frilly edge curling probability prediction model and the frilly edge curling occurrence label.

[0188] The initial model for predicting the probability of edge curling of the lotus leaf includes:

[0189] A first long short-term memory neural network is used to construct a physical driving channel. The physical driving channel takes the edge-middle water content gradient sequence as input, extracts the water content gradient temporal cumulative features, and outputs a physical driving feature vector.

[0190] A second long short-term memory neural network is used to construct a hysteresis correlation channel. The hysteresis correlation channel takes the humidity shock index sequence as input, extracts the hysteresis response features from humidity shock to moisture content gradient change, and outputs a hysteresis correlation feature vector.

[0191] A third long short-term memory neural network is used to construct a regulation effect channel. The regulation effect channel takes the thermal shock index sequence as input, extracts the inverse regulation features of thermal shock on water content gradient, and outputs a regulation effect feature vector.

[0192] A probabilistic fusion output layer is constructed. The probabilistic fusion output layer concatenates the physical driving feature vector, the hysteresis correlation feature vector, and the moderating effect feature vector in the feature dimension to obtain a concatenated fusion vector. The concatenated fusion vector is mapped to a one-dimensional feature value through a fully connected layer. Then, the one-dimensional feature value is mapped to a probability value of the ruffled edge curling occurring in a future preset time period through a Sigmoid activation function.

[0193] The physical driving channel, the hysteresis correlation channel, the modulation effect channel, and the probability fusion output layer are connected to form an initial lotus leaf edge curling probability prediction model.

[0194] The differential pressure calculation and execution module 14 is used to calculate the differential target pressure value at both ends of the smoothing pressure roller based on the edge-middle moisture content gradient, the paper edge warping height and the paper weight when the probability value is greater than or equal to a preset probability threshold, and control the execution pressure of the smoothing pressure roller according to the differential target pressure value.

[0195] Specifically, when the probability value is greater than or equal to a preset probability threshold, the differential target pressure values ​​at both ends of the smoothing roller are calculated based on the edge-to-center moisture content gradient, the paper edge warping height, and the paper basis weight. The execution pressure of the smoothing roller is then controlled according to these differential target pressure values, including:

[0196] The probability value is compared with the preset probability threshold. When the probability value is greater than or equal to the preset probability threshold, the edge-to-middle moisture content gradient, the paper edge warping height and the paper weight of the current batch of printed label paper are input into the pre-trained smoothing pressure calculation model. The smoothing pressure calculation model outputs the predicted differential target pressure value.

[0197] The bias correction amount is determined based on the historical prediction error direction distribution of the smoothing pressure calculation model on the validation set;

[0198] The predicted differential target pressure value is added to the bias correction amount to obtain the differential target pressure value at both ends of the smoothing roller;

[0199] According to the differentiated target pressure value, the output pressure of the electric proportional valves of the cylinders at both ends of the smoothing roller is controlled respectively, so that different execution pressures are applied to the two ends of the smoothing roller.

[0200] The pre-training process of the stress relief calculation model includes:

[0201] Obtain a historical control sample set. Each historical control sample in the historical control sample set contains an edge-to-middle moisture content gradient, paper edge warping height, and paper weight in a historical edge curling control event, as well as a corresponding optimal differentiated target pressure value label.

[0202] The historical regulation sample set is augmented to obtain an augmented historical regulation sample set.

[0203] An initial stress calculation model was constructed based on a fully connected neural network.

[0204] Using the edge-to-middle moisture content gradient, paper edge warping height, and paper basis weight in the enhanced historical control sample set as input features, and the corresponding optimal differentiated target pressure value label as output label, the initial smoothing pressure calculation model is trained in a supervised manner until the verification convergence is obtained, thus obtaining the trained smoothing pressure calculation model.

[0205] Specifically, the historical regulation sample set is augmented to obtain an augmented historical regulation sample set, which includes:

[0206] For each historical control sample in the historical control sample set, obtain the edge-to-middle moisture content gradient and the paper edge warping height of the historical control sample;

[0207] Within a preset perturbation range, a first random micro-perturbation is applied to the edge-middle moisture content gradient of the historical control sample, and a second random micro-perturbation is applied to the paper edge warping height of the historical control sample, thereby generating the perturbed edge-middle moisture content gradient and the perturbed paper edge warping height.

[0208] The perturbed edge-to-middle moisture content gradient and the perturbed paper edge warp height are combined with the paper weight of the historical control sample to form the input features of the enhanced sample. The optimal differentiated target pressure value label corresponding to the historical control sample is used as the output label of the enhanced sample.

[0209] The enhanced samples generated from all historical control samples are merged with the historical control sample set to obtain the enhanced historical control sample set.

[0210] Based on the historical prediction error direction distribution of the smoothing pressure calculation model on the validation set, the bias correction amount is determined, including:

[0211] Construct a validation set containing multiple validation samples, where each validation sample includes the edge-to-center moisture content gradient, paper edge warping height, paper basis weight, and the corresponding optimal differentiated target pressure value label;

[0212] For each validation sample in the validation set, the edge-to-center moisture content gradient of the validation sample, the paper edge warping height of the validation sample, and the paper weight are input into the trained smoothing pressure calculation model, and the predicted differential target pressure value of the validation sample is output.

[0213] For each validation sample in the validation set, the optimal differentiated target pressure value label of the validation sample is calculated, and the predicted differentiated target pressure value of the validation sample is subtracted to obtain the prediction error value of the validation sample.

[0214] The arithmetic mean of the prediction error values ​​of all validation samples in the validation set is used as the bias correction.

[0215] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0216] This specification and accompanying drawings are merely illustrative examples of the invention and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its scope. Therefore, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is intended to include these modifications and modifications.

Claims

1. A paper smoothing pressure adaptive control method based on intelligent manufacturing, characterized in that, include: Collect data on the moisture content at the edge of the paper, the moisture content in the middle of the paper, the humidity of the workshop environment, the printing speed, the drying temperature, and the warping height at the edge of the paper on the label printing production line, and obtain the paper weight of the current batch of printed label paper. The edge-to-middle moisture content gradient is calculated based on the moisture content at the edge of the paper and the moisture content in the middle of the paper. The rate of change of ambient humidity is calculated based on the ambient humidity at multiple consecutive sampling times. The thermal shock index is calculated based on the drying temperature and the printing speed. The humidity shock index is calculated based on the rate of change of ambient humidity and the paper weight of the current batch of printed label paper. The pre-trained frilly edge curling probability prediction model outputs the probability value of frilly edge curling within a preset time period based on the edge-middle moisture content gradient, the humidity shock index, and the thermal shock index. When the probability value is greater than or equal to the preset probability threshold, the differential target pressure value at both ends of the smoothing roller is calculated based on the edge-middle moisture content gradient, the paper edge warping height and the paper weight, and the execution pressure of the smoothing roller is controlled according to the differential target pressure value.

2. The paper smoothing pressure adaptive control method based on intelligent manufacturing according to claim 1, characterized in that, Collect data on the moisture content at the paper edges, the moisture content in the middle of the paper, the ambient humidity in the workshop, the printing speed, the drying temperature, and the edge warping height of the paper on the label printing production line, and obtain the paper weight of the current batch of printed labels, including: Between the unwinding station and the smoothing roller in the printing label paper production line, the moisture content of the paper edges is collected by infrared moisture meters deployed at both edges in the paper width direction, and the moisture content of the paper middle is collected by infrared moisture meters deployed in the middle of the paper width direction. Temperature and humidity sensors are deployed in the workshop where the printing label paper production line is located to collect the ambient temperature and humidity of the workshop. The drying temperature is collected by temperature sensors deployed at the drying station. The paper edge warping height is collected by laser displacement sensors deployed at both edges along the width of the paper. Get the current printing speed and the paper weight of the current batch of printed labels.

3. The paper smoothing pressure adaptive control method based on intelligent manufacturing according to claim 1, characterized in that, The edge-to-center moisture content gradient is calculated based on the paper edge moisture content and the paper center moisture content. The environmental humidity change rate is calculated based on the workshop environmental humidity at multiple consecutive sampling times. The thermal shock index is calculated based on the drying temperature and the printing speed. The humidity shock index is calculated based on the environmental humidity change rate and the paper weight of the current batch of printed label paper, including: Subtracting the paper's middle moisture content from the paper's edge moisture content at the same sampling time yields the edge-middle moisture content gradient at the same sampling time. Subtract the humidity of the workshop environment at the previous sampling time from the humidity at the current sampling time to obtain the humidity difference. Divide the humidity difference by the time interval between the current sampling time and the previous sampling time to obtain the rate of change of the ambient humidity at the current sampling time. The thermal shock index is obtained by dividing the difference between the collected drying temperature and the workshop ambient temperature at the same sampling time by the collected printing speed. The humidity shock index is calculated by multiplying the rate of change of ambient humidity at the current sampling time by the paper weight of the current batch of printed label paper.

4. The paper smoothing pressure adaptive control method based on intelligent manufacturing according to claim 1, characterized in that, Using a pre-trained frilly edge curling probability prediction model, based on the edge-to-center moisture content gradient, the humidity shock index, and the thermal shock index, the model outputs the probability value of frilly edges curling within a preset time period in the future, including: The edge-intermediate moisture content gradients of multiple consecutive sampling times before the current sampling time are used to form an edge-intermediate moisture content gradient sequence, and the humidity shock indexes of multiple consecutive sampling times before the current sampling time are used to form a humidity shock index sequence. The edge-to-middle moisture content gradient sequence, the humidity shock index sequence, and the thermal shock index sequence are input into a pre-trained frilly edge curling probability prediction model. The physical driving channel of the frilly edge curling probability prediction model outputs a physical driving feature vector based on the edge-to-middle moisture content gradient sequence. The hysteresis correlation channel of the frilly edge curling probability prediction model outputs a hysteresis correlation feature vector based on the humidity shock index sequence. The moderating effect channel of the frilly edge curling probability prediction model outputs a moderating effect feature vector based on the thermal shock index sequence. Then, the probability fusion output layer of the frilly edge curling probability prediction model fuses the physical driving feature vector, the hysteresis correlation feature vector, and the moderating effect feature vector to output the probability value of frilly edge curling occurring within a preset time period in the future.

5. The paper smoothing pressure adaptive control method based on intelligent manufacturing according to claim 1, characterized in that, The pre-training process of the ruffle edge curling probability prediction model includes: Obtain a historical production sample set. Each historical production sample in the historical production sample set contains an edge-intermediate moisture content gradient sequence, a humidity shock index sequence, and a thermal shock index sequence for multiple consecutive sampling times in a historical production batch, as well as a corresponding fringe curling occurrence label. The fringe curling occurrence label indicates whether fringe curling actually occurred in the future preset time period in the historical production batch. An initial fringe edge curling probability prediction model is constructed, which includes a physical driving channel, a hysteresis correlation channel, a moderating effect channel, and a probability fusion output layer. The edge-to-intermediate moisture content gradient sequence, humidity shock index sequence, and thermal shock index sequence from the historical production sample set are used as input features. The corresponding frilly edge curling occurrence labels are used as output labels. The training objective is to minimize the binary cross-entropy loss function. The initial frilly edge curling probability prediction model is trained in a supervised manner until it is verified to converge, thus obtaining the trained frilly edge curling probability prediction model. The binary cross-entropy loss function calculates the error between the probability value output by the initial frilly edge curling probability prediction model and the frilly edge curling occurrence label.

6. The paper smoothing pressure adaptive control method based on intelligent manufacturing according to claim 5, characterized in that, Construct an initial prediction model for the probability of edge curling of the ruffled leaf, including: A first long short-term memory neural network is used to construct a physical driving channel. The physical driving channel takes the edge-middle water content gradient sequence as input, extracts the water content gradient temporal cumulative features, and outputs a physical driving feature vector. A second long short-term memory neural network is used to construct a hysteresis correlation channel. The hysteresis correlation channel takes the humidity shock index sequence as input, extracts the hysteresis response features from humidity shock to moisture content gradient change, and outputs a hysteresis correlation feature vector. A third long short-term memory neural network is used to construct a regulation effect channel. The regulation effect channel takes the thermal shock index sequence as input, extracts the inverse regulation features of thermal shock on water content gradient, and outputs a regulation effect feature vector. A probabilistic fusion output layer is constructed. The probabilistic fusion output layer concatenates the physical driving feature vector, the hysteresis correlation feature vector, and the moderating effect feature vector in the feature dimension to obtain a concatenated fusion vector. The concatenated fusion vector is mapped to a one-dimensional feature value through a fully connected layer. Then, the one-dimensional feature value is mapped to a probability value of the ruffled edge curling occurring in a future preset time period through a Sigmoid activation function. The physical driving channel, the hysteresis correlation channel, the modulation effect channel, and the probability fusion output layer are connected to form an initial lotus leaf edge curling probability prediction model.

7. The paper smoothing pressure adaptive control method based on intelligent manufacturing according to claim 1, characterized in that, When the probability value is greater than or equal to a preset probability threshold, the differential target pressure values ​​at both ends of the smoothing roller are calculated based on the edge-to-center moisture content gradient, the paper edge warping height, and the paper basis weight. The execution pressure of the smoothing roller is then controlled according to these differential target pressure values, including: The probability value is compared with the preset probability threshold. When the probability value is greater than or equal to the preset probability threshold, the edge-to-middle moisture content gradient, the paper edge warping height and the paper weight of the current batch of printed label paper are input into the pre-trained smoothing pressure calculation model. The smoothing pressure calculation model outputs the predicted differential target pressure value. The bias correction amount is determined based on the historical prediction error direction distribution of the smoothing pressure calculation model on the validation set; The predicted differential target pressure value is added to the bias correction amount to obtain the differential target pressure value at both ends of the smoothing roller; According to the differentiated target pressure value, the output pressure of the electric proportional valves of the cylinders at both ends of the smoothing roller is controlled respectively, so that different execution pressures are applied to the two ends of the smoothing roller.

8. The paper smoothing pressure adaptive control method based on intelligent manufacturing according to claim 7, characterized in that, The pre-training process of the stress relief calculation model includes: Obtain a historical control sample set. Each historical control sample in the historical control sample set contains an edge-to-middle moisture content gradient, paper edge warping height, and paper weight in a historical edge curling control event, as well as a corresponding optimal differentiated target pressure value label. The historical regulation sample set is augmented to obtain an augmented historical regulation sample set. An initial stress calculation model was constructed based on a fully connected neural network. Using the edge-to-middle moisture content gradient, paper edge warping height, and paper basis weight in the enhanced historical control sample set as input features, and the corresponding optimal differentiated target pressure value label as output label, the initial smoothing pressure calculation model is trained in a supervised manner until the verification convergence is obtained, thus obtaining the trained smoothing pressure calculation model. Specifically, the historical regulation sample set is augmented to obtain an augmented historical regulation sample set, which includes: For each historical control sample in the historical control sample set, obtain the edge-to-middle moisture content gradient and the paper edge warping height of the historical control sample; Within a preset perturbation range, a first random micro-perturbation is applied to the edge-middle moisture content gradient of the historical control sample, and a second random micro-perturbation is applied to the paper edge warping height of the historical control sample, thereby generating the perturbed edge-middle moisture content gradient and the perturbed paper edge warping height. The perturbed edge-to-middle moisture content gradient and the perturbed paper edge warp height are combined with the paper weight of the historical control sample to form the input features of the enhanced sample. The optimal differentiated target pressure value label corresponding to the historical control sample is used as the output label of the enhanced sample. The enhanced samples generated from all historical control samples are merged with the historical control sample set to obtain the enhanced historical control sample set.

9. The paper smoothing pressure adaptive control method based on intelligent manufacturing according to claim 7, characterized in that, Based on the historical prediction error direction distribution of the smoothing pressure calculation model on the validation set, the bias correction amount is determined, including: Construct a validation set containing multiple validation samples, where each validation sample includes the edge-to-center moisture content gradient, paper edge warping height, paper basis weight, and the corresponding optimal differentiated target pressure value label; For each validation sample in the validation set, the edge-to-center moisture content gradient of the validation sample, the paper edge warping height of the validation sample, and the paper weight are input into the trained smoothing pressure calculation model, and the predicted differential target pressure value of the validation sample is output. For each validation sample in the validation set, the optimal differentiated target pressure value label of the validation sample is calculated, and the predicted differentiated target pressure value of the validation sample is subtracted to obtain the prediction error value of the validation sample. The arithmetic mean of the prediction error values ​​of all validation samples in the validation set is used as the bias correction.

10. A paper smoothing pressure adaptive control system based on intelligent manufacturing, characterized in that, The paper smoothing pressure adaptive control method based on intelligent manufacturing as described in any one of claims 1-9 includes: The multi-source data acquisition and preprocessing module is used to collect data on the moisture content of paper edges, the moisture content of paper center, the humidity of the workshop environment, the printing speed, the drying temperature, and the warping height of paper edges on the printing label paper production line, and to obtain the paper weight of the current batch of printed label paper. The feature index generation module is used to calculate the edge-to-middle moisture content gradient based on the moisture content of the paper edge and the moisture content of the paper middle, calculate the environmental humidity change rate based on the workshop environmental humidity at multiple consecutive sampling times, calculate the thermal shock index based on the drying temperature and the printing speed, and calculate the humidity shock index based on the environmental humidity change rate and the paper weight of the current batch of printed label paper. The lotus leaf edge curling probability prediction module is used to output the probability value of lotus leaf edge curling within a preset time period in the future, based on the edge-middle moisture content gradient, the humidity impact index and the thermal shock index, using a pre-trained lotus leaf edge curling probability prediction model. The differential pressure calculation and execution module is used to calculate the differential target pressure value at both ends of the smoothing pressure roller based on the edge-middle moisture content gradient, the paper edge warping height and the paper weight when the probability value is greater than or equal to a preset probability threshold, and control the execution pressure of the smoothing pressure roller according to the differential target pressure value.

Images